Maximizing Signal Integrity: Fine-Tuning Via Impedance In HDFO Architectures

The most different aspect between a normal lamination structure and High-Density Fan-out (HDFO) is the routing scale. That aspect is also the challenge and focus of this study. At an HDFO scale, most of the electrical properties cannot be measured by instruments. Therefore, this study uses the indirect method to determine the impedance information of the via and match the impedance. Since the via structure is very complex and varies, this study will focus on parameters such as the metal part and single-end channel to reduce the variables. It uses the design of experiments (DOE) method to compare the different combinations. Finally, it will provide optimized results and the changed curve.

Methodology

The analysis is separated into two parts. The first part is in the frequency domain, using insertion loss and return loss to interpret whether the impedance is matched or not. The second part uses de-embed technology to extend the channel length to emphasize the jitter/overshot, etc.

Fig. 1: HDFO via model on the electronic design automation (EDA) tool.

Figure 1 is the fundamental model for the design of experiments (DOEs). This study limits the variables into three items: via length, via radius and via anti-pad. To keep the trace impedance as accurate as 50 ohms, a four-layer lamination structure was chosen, the trace width was fixed at 2 µm, and the dielectric constant of epoxy, D_k, was set at 3.1. Since this structure was already verified in Amkor products, the characteristic impedance in segments of top and bottom traces can be treated as the ideal 50 ohms.

Fig. 2: Different via radius TDR results.

Figure 2 shows the reason why time domain reflectometry (TDR) cannot work in the HDFO scale. Since the minimum impulse wavelength is too large for the channel length, the reflection methodology cannot perform the HDFO via structure analysis as a printed circuit board (PCB) via. In the general case, the via always shows the low impedance and valley-shape curve in the TDR results. Enlarging the via properties and impedance information are the most challenging aspects in this study.

Results and data analysis

This study uses three DOE groups and both frequency and time domain indices to determine the impact of variable changes. Finally, this process will determine the key factors and introduce a methodology to optimize the via impedance in the early design stage.

Table 1: DOE groups and changing variable.

In the frequency domain analysis, the differences between each variable change are very small. Thus, determining the valid variable(s) is the most difficult part in frequency domain analysis. Because each DOE’s performance shows different features, the measurement point cannot be set at the fixed frequency point. The measurement point must adhere to the following conditions. (1) The return loss should be near the DC point; (2) the first peak should be between (15~20 GHz); (3) the first valley should be between (40~45 GHz); and the insertion loss should always be a linear curve, so the measurement point will follow the DOE’s return loss (see Figure 3).

Fig. 3: Measurement frequency point example. (DOE group1.)

For the next step, the measurement data takes the division of min (return loss)/ max (insertion loss) value because the absolute value of insertion loss or return loss is very small. Using the scale as a ratio to evaluate the difference increases the value and adjusts the results to the same level. This makes analyzing the data easier when compared with other DOE groups. DOE results are shown in Figures 4 to 9.

Fig. 4: DOE group1. Ratio by different via radius.

Fig. 5: DOE group2. Ratio by different PI layer thickness.

Fig. 6: DOE group2. Ratio by different via anti pad radius.

Obviously, the difference between variable changes is very small, even after normalizing. That confirms the challenging aspect of this study. Since the frequency domain cannot provide the decisive factor of via performance, using the time domain method to determine the via quality by eye diagram properties could be preferred. The next step is finding the best eye diagram in the time domain and comparing it with frequency domain data to verify which approach can provide the best results.

For the time domain analysis, this study used the ideal signal source to create a 30 Gbps signal for eye diagram simulation and chose four measurement indices to judge the eye diagram quality: peak to peak eye jitter, RMS eye jitter, eye opening factor and eye signal to noise ratio. The comprehensive index can catch the tiny differences which cannot be directly measured. For clarity, each ratio type index was divided by the minimum value to provide one point as the reference for normalization in each group. The algorithm from the normalization procedure converts the different indices into the same scale.

Fig. 7: Group1. Eye diagram index.

Fig. 8: Group1. Eye diagram index after using algorithm.

In DOE Group1, each via radius, 3 µm, 3.5 µm, and 4 µm looks great, but the best one cannot be determined. However, by using the standard deviation of the data, a clear, relative relationship is obtained, as Figure 9 shows.

Fig. 9: Group1. Index after taking standard deviation.

In DOE Group1, the relationship of via radii, 3 µm and 3.5 µm shows their improved performance. Therefore, all steps of the data analysis process are complete. More results are shown in Figures 10 and 11.

Fig. 10: Group2. Data analysis results.

Fig. 11: Group3. Data analysis results.

Conclusion

The goal of this research is to create a method to optimize via performance at the High-Density Fan-Out scale. When the impedance value cannot be determined, a high-performance evaluation of the signal transmission provides a viable alternative. By applying this approach to products in their early design stage, it can help to verify the via performance and reduce the risk of mistakes.

The summary can be separated into two parts. First, frequency domain cannot show the differences directly. When the time domain method gets improved eye diagram results of the via structure, the important features are revealed. With good time domain results based on the standard deviation, taking the standard deviation on frequency domain data provides additional information. The results are shown in Figures 12 to 14.

Fig. 12: Group1. Standard deviation comparison.

Fig. 13: Group2. Standard deviation comparison.

Fig. 14: Group3. Standard deviation comparison.

Based on these results, the evaluation of a good via structure provides two rules.

First, the standard deviation value of these three indices will be very close. As Figure 14 shows, when the time domain results are used to judge, 7 µm and 13 µm are at the same level. Adding frequency domain data to check the results changes that decision. Although return loss and insertion loss values seem unrelated, using the standard deviation approach indicates that the 7 µm case matches the required conditions. As a result, the better via structure emerged from these two candidates.

Secondly, the measurement point is important. Near DC/Peak/Valley locations not only indicate the low/middle/high frequency, but also include the changing point. At these edge points, the channel response will be enlarged and help to recognize the subtle differences.

Tiny vias that ignore the mismatching effect account for less than 5% of the entire channel. However, in the package industry, IC functions are becoming smaller, faster, and more complex. As a result, guaranteeing the signal integrity of the package becomes even more important. Reducing the risk and development cycle time is the key to successful product introduction. Also, improving even the smallest detail is the best way to increase the product’s quality.

References

1. Bogatin, Eric, Signal and Power Integrity — Simplified, Pearson Education, 2010.
2. S. G. Hsu and R. B. Wu, “Full-wave characterization of a through hole via in multi-layered packaging,” IEEE Trans. Microwave Theory Tech., vol. 43, pp. 1073-1081, May 1995.
3. John Lee and Mike Kelly, “Amkor’s 2.5D Package and HDFO – Advanced Heterogeneous Packaging Solutions,” 2018.
4. Jiang, X., et al., “Designing 224G High performance FPGA package and board with confidence,” DesignCon, USA, 2021.
5. Xing, W., et al., “Optimizing BGA Ball Pattern for Signal Integrity,” DesignCon, USA, 2021.